spark as the gateway drug to typed functional programming: spark summit east talk by jeff smith and...

Spark as the Gateway Drug To Typed Functional Programming

Jeff Smith Rohan Aletty x.ai

Real World AI• Scale is increasing • Complexity is increasing • Human brain size is constant

System Complexity

Data Ingest Annotation Routing

Response Generation

Annotation Services

Models

Annotation Services

Models

Annotation Services

Models

Annotation Services

Models

Annotation Services

Models

Annotation Services

Models

Annotation Services

Models

Models

Annotation Services

Knowledge Base

Problem Complexity

Complex Intelligence

Datanauts

Scala• Bleeding edge • Real world

Spark• Incredibly powerful • Easy to use

Typed Functional Programming• Powerful abstractions • Tough learning curve

Functions

Methods• Collection of statements • Might have side effects • On an object

Methodspublic class Dataset {

private List<Double> observations; private Double average;

public Dataset(List<Double> inputData) { observations = inputData; }

}


public double getAverage() { Double runningSum = 0.0;

for (Double observation : observations) { runningSum += observation; }

average = runningSum / observations.size();

return average; }}


public void setObservations(List<Double> inputData) { observations = inputData; }}


private List<Double> observations; private Double average;

public Dataset(List<Double> inputData) { observations = inputData; }

public double getAverage() { Double runningSum = 0.0;

for (Double observation : observations) { runningSum += observation; }

average = runningSum / observations.size();

return average; }

public void setObservations(List<Double> inputData) { observations = inputData; }}

Functions• Collection of expressions • Returns a value • Are objects (in Scala) • Can be in-lined

Functions in Scalaval inputData = List(1.0, 2.0, 3.0)

Functions in Scaladef average(observations: List[Double]) { observations.sum / observations.size}

average(inputData)

Functions in Scaladef add(x: Double, y: Double) = { x + y}

val sum = inputData.foldLeft(0.0)(add)

val average = sum / inputData.size

Functions in Scalaval sum = inputData.foldLeft(0.0)(add)

val average = sum / inputData.size

inputData.foldLeft(0.0)(_ + _) / inputData.size

Functions in SparkinputData.foldLeft(0.0)(_ + _) / inputData.size

val observations = sc.parallelize(inputData)

observations.fold(0.0)(_ + _) / observations.count()

Immutability

Mutation• Changing an object

Mutationvisits = {"Church": 2, "Backus": 1, "McCarthy": 4}

old_value = visits["Backus"]

visits["Backus"] = old_value + 1

Immutability• Never changing objects

Immutability in Scalaval visits = Map("Church" -> 2, "Backus" -> 1, "McCarthy" -> 4)

val updatedVisits = visits.updated("Backus", 2)

Immutability in Sparkval manyVisits = sc.parallelize(visits.toSeq)

val additionalVisit = sc.parallelize(Seq(("Backus", 1)))

val updatedVisits = manyVisits.union(additionalVisit) .aggregateByKey(0)(_ + _, _ + _)

Concepts• Higher-order functions • Anonymous functions • Purity of functions

Concepts• Currying • Referential transparency • Closures • Resilient Distributed Datasets

Lazy Evaluation

Functional Programming — Lazy Evaluation

• Delaying evaluation of an expression until a value is needed

• Two major advantages of lazy evaluation • Deferring computation allows program only evaluate what is necessary • Changing evaluation scheme into to be more efficient

Spark — Lazy Evaluation• All transformations are lazy

• Their existence added to Spark computation DAG

• Example DAGs

Spark — Lazy Evaluationval rdd1 = sc.parallelize(...)

val rdd2 = rdd1.map(...)



rdd3.take(5)

Spark — Learning Laziness• Advantage 1: (deferred computation)

• draws directly from only evaluating parts of DAG that are necessary when executing an action

• Advantage 2: (optimized evaluation scheme) • draws directly from pipelining within Spark stages to make execution

more efficient

Functional Programming — Type Systems

• Mechanism for defining algebraic data types (ADTs) which are useful for program structure • i.e. “let’s group this data together and brand it a new type”

• Compile time guarantees of correctness of program • e.g. “no, you cannot add Foo to Bar”

Spark — Types• RDD’s (typed), Datasets (typed), DataFrames (untyped)

• Types provide great schema enforcement on a dataset for preventing unexpected behavior

Spark — Typescase class Person(name: String, age: Int)

val peopleDS = spark.read.json(path).as[Person]

val ageGroupedDs = peopleDS.groupBy(_.age)

Spark — Learning Types• Spark through Scala also allows learning of pattern

matching • ADTs as both product types and union types

• Allows us to reason about code easier

• Gives us compile time safety

Spark — Learning Typestrait Person { def name: String }

case class Student(name: String, grade: String) extends Person

case class Professional(name: String, job: String) extends Person

val personRDD: RDD[Person] = sc.parallelize(…)

// working with both union and product typesval mappedRDD: RDD[String] = personRDD.map { case Student(name, grade) => grade case Professional(name, job) => job}

Spark — Learning Typesval rdd1: RDD[Person] = sc.parallelize(...)

val rdd2: RDD[String] = rdd1.map("name: " + _) // Compilation error!

val rdd3: RDD[String] = rdd2.map("name: " + _.name) // It works!

Monads

Functional Programming — Monads• In category theory:

• “a monad in X is just a monoid in the category of endofunctors of X”

• In functional programming, refers to a container that can: • Inject a value into the container • Perform operations on values returning a container with new values • Flatten nested containers into a single container

Scala — Monads!trait Monad[M[]] { // constructs a Monad instance from the given value, e.g. List(1) def apply[T](v: T): M[T]

// effectively lets you transform values within a Monad def bind[T, U](m: M[T])(fn: (T) => M[U]): M[U]}

Scala — Monads!• Many monads in Scala

• List, Set, Option, etc.

• Powerful line of thinking • Helps code comprehension • Reduces error checking logic (pattern matching!) • Can build further transformations: map(), filter(), foreach(), etc.

Spark — Learning Monads?• We have many “computation builders” -- (RDD’s, Datasets,

DataFrames) • Containers on which transformations can be applied

• Similar to monads though not identical • No unit function to wrap constituent values • Cannot lift all types into flatMap function unconstrained

For Later

Conclusions• Spark introduces all types of devs to Scala

• Scala helps people learn typed functional programming

• Typed functional programming improves Spark development

x.ai @xdotai [email protected] New York, New York

Use the code ctwsparks17 for 40% off!

Thank You

spark as the gateway drug to typed functional programming: spark summit east talk by jeff smith and...

Data & Analytics